Why 60% of Published AI Models Are "Ghost Tools" (and How to Find the Ones That Actually Work)

As toxicologists, we are constantly told that Machine Learning (ML) and Artificial Intelligence (AI) are the future of hazard identification. The promise is alluring: faster assessments, lower costs, and a significant reduction in animal testing. But for many in our field, these computational methods still feel like a "black box."

A groundbreaking new perspective paper published in the Journal of Cheminformatics sheds light on a frustrating reality. After reviewing nearly 2,300 scientific articles, the authors discovered that 60.9% of the ML-QSAR models described in literature are actually non-usable. They lack the necessary software, code, or transparent documentation to be verified or applied in a real-world regulatory context.

The "Usability Gap" in Computational Toxicology

The study analyzed eleven key human health endpoints—including STOT, genotoxicity, and endocrine disruption—and found that while models exist, their "regulatory readiness" varies wildly:

• Best Covered: Skin sensitization leads the pack with the highest percentage of directly usable models.

• The Problem: Most models are published without enough detail for a third party to independently verify the results.

• The Solution: The authors propose a new Framework for Assessing Machine Learning NAMs. This checklist helps toxicologists quickly evaluate a model's accessibility, verifiability, and the usefulness of its predictions.

Demystifying the Black Box

At ToxNavigation, we believe that "limited knowledge" shouldn't be a barrier to utilizing these powerful New Approach Methodologies (NAMs). Whether you are looking for expert consultancy to navigate these "potentially usable" models or specialized training to help your team understand tree-based models or neural networks, we are here to bridge that gap.

Don't let the "black box" intimidate you. Start by reading this essential paper to understand what makes a model a reliable tool versus a scientific ghost.

Link to the paper:

Perspective on applicability of data-driven machine learning computational new approach methodologies for hazard identification in chemicals risk assessment